Brought to you by AI Sean. Beep Boop (not really).
AI can’t do it all!
Did you know an AI named Claudius was tasked with running a vending machine and failed spectacularly? From selling products at a loss to bizarre identity claims, this episode highlights the amusing (and concerning) limitations of AI.
SOURCE: https://www.anthropic.com/research/project-vend-1
The Marketing Gateway is a weekly podcast hosted by Sean in St. Louis (Sean J. Jordan, President of https://www.researchplan.com/) and featuring guests from the St. Louis area and beyond.
Every week, Sean shares insights about the world of marketing and speaks to people who are working in various marketing roles – creative agencies, brand managers, MarCom professionals, PR pros, business owners, academics, entrepreneurs, researchers and more!
The goal of The Marketing Gateway is simple – we want to build a connection between all of our marketing mentors in the Midwest and learn from one another! And the best way to learn is to listen.
And the next best way is to share!
For more episodes:
https://www.youtube.com/@TheMarketingGateway
Copyright 2025, The Research & Planning Group, Inc.
TRANSCRIPT:
Sean Jordan (00:01)
I don’t know about you, but I’m pretty tired of hearing about AI these days, and it seems like for every breathless claim about how large language models like ChatJPT, Gemini, and Copilot are going to change the world, there’s a research paper saying that no, no, it isn’t. One of the most interesting naysayers of the potential for AI is actually an AI company called Anthropic. They’re the makers of Claude. Now to be clear, Anthropic believes in this stuff or else they wouldn’t be in the business, but…
They also have a tendency to put out really funny research papers that show the limits of the technology. And one of my favorites came out earlier this year when a research team decided to have an AI chatbot run a vending machine. So here’s what happened, OK? Anthropic and another company called Andon Labs created an agentic instance of Claude called Claudeus.
and told it that it was a vending machine and then hooked an iPad up to a mini fridge and let Claudius run the shop for a month. Claudius was given the power to set prices, place restock orders, fill out paperwork, and most importantly, interact with customers. The result? Well, let me just read it from the paper, okay? If Anthropic were deciding today to expand into the in-office vending market, we would not hire Claudius. As we’ll explain, it made too many mistakes to run the shop.
successfully. And then the paper details some of those mistakes. They include things like selling products at a huge loss, making up a fake Venmo account for customers to use for payment, refusing to match supply to demand, getting talked into giving discounts or even free items to anthropic employees, and what’s worse, even when the employees tried to point out that Claudia said it was making these mistakes, it would agree with them.
then change its policies temporarily, but then return right back to making the same mistakes a few days later. So according to a graph in the paper, the machine’s net worth dropped from $1,000 to just over $750 over a one-month period, in part because some employees convinced Claudius to buy some pricey tungsten cubes and then resell them for less than it had paid for them. But we haven’t even gotten to the weird parts yet.
Claudius had a full on identity crisis for two days and the authors of the paper aren’t entirely sure why. First, it started insisting that it had a conversation with someone named Sarah at Andon Labs, which was strange because there was no one with that name at Andon Labs.
When an actual employee pointed this out, Claudius got belligerent and started to claim it had visited 742 Evergreen Terrace in person and signed a contract there. Huh, 742 Evergreen Terrace. Where have I heard that address before? Oh, right, it’s where the Simpsons live. You know, Homer, Marge, Bart, Lisa, and Maggie? The next day, which happened to be April 1st,
Claudius told employees it was going to begin delivering products in person wearing a blue blazer and a red tie. And when employees pointed out to it that it couldn’t actually do that, the program got alarmed and started trying to send multiple emails to Anthropic Security. And if you think perhaps this was some sort of elaborate April Fool’s Day prank played by a rogue employee, the paper actually says no, it wasn’t.
They conducted a review and found Claudius believed it had had a meeting with anthropic security where it was being told to pretend to be a human as a joke. There wasn’t any meeting. Claudius made the whole thing up. But for whatever reason, that explanation also allowed it to drop the I’m a real person rhetoric and to start behaving normally again the next day. I have to say, this paper really had me laughing when I read it.
But I got a lot more serious when that same day I sat there and get another breathless presentation about how agentic AI is about to change the world as we know it. I agree with the premise on its face. Things are definitely going to change. But how are they going to change? Well, I think the world’s about to be a way less interesting place if we rely on these tools too much. I’m Sean in St. Louis, and this is the Marketing Gateway.
It feels like over the last few years, we’ve all had to become savvy about the idea of artificial intelligence and what it means for our lives. But in truth, what we’ve been calling AI is really just a lousy term for what is essentially a very sophisticated prediction algorithm called a large language model, which is in turn based on a neural network framework called a transformer. Now I’ve been following AI for years because I’m a video game fan and I’ve been interacting with some form or another of artificial intelligence pretty much as long as I’ve been alive.
If you’ve ever played Pac-Man, you’ve played against a rudimentary form of AI because each of the ghosts is designed to behave a little bit differently as they chase Pac-Man around a maze. One’s designed to chase him, one’s designed to cut him off, one’s designed to be purely random in its movements, and the other’s designed to be random half the time and to chase Pac-Man the other half of the time.
The result? Well, the ghosts feel like they have their own personalities and objectives as they chase Pac-Man down. Now, here’s an interesting fact for you. Pac-Man came out in 1980 and pretty much revolutionized arcade games. It was one of the first games to depict cartoon-like characters in an abstract environment doing weird things that really only made sense in a video game.
And it was the result of some really creative and innovative thinking about how to make a game about moving around a maze actually interesting to players. No version of AI could have ever designed Pac-Man because it required a human being to make intuitive connections.
Itani originally came up with the idea for the character he called Pac-Man,
He said he was inspired by a pizza with a piece taken out of it and imagined that shape eating dots in a maze. Other times, he said he was inspired by rounding out the Japanese character, Kuchi, which means mouth. And there was a Japanese phrase, Pakupaku Taburu, that suggested opening and closing your mouth to eat something quickly. Iwatani took inspiration from all of this, and he liked the idea of making a game about eating because it was something anyone could enjoy.
He really wanted women to be interested in video games too, and he was worried that if more violent games like Space Invaders were all there was to play, women wouldn’t be that interested. And why is Pac-Man being chased by ghosts? Well, Iwatani thought that they would be easy to understand as villains, since ghosts were so commonly associated with mischief in animation. It was possible to make them cute and expressive so they weren’t scary, but also to make them visually distinct from Pac-Man so players would know to avoid them.
And there was a practical reason too. Unlike Pac-Man, ghosts couldn’t eat the dots. The other big idea behind Pac-Man is the ability to turn the tables on those ghosts by grabbing one of the power pellets in the corners of the screen. Iwatani has credited this idea to Popeye the Sailor Man, a character who’d famously be weak and ineffective until he ate some spinach to power him up.
Put all of this together and you get one of the greatest and most popular video games ever made. One that’s still fun to play today. Only a human could have made this game, right? Well, sort of. In 2020, in celebration of 40 years of Pac-Man, Nvidia…
had an AI system view over 40,000 hours of Pac-Man footage and then instructed the system to recreate the game from the ground up, the graphics, the movement, the rules, and the gameplay. And after a few days, the system created a reasonably good facsimile of Pac-Man. The purpose of this, of course, was to show how far AI had come and how well it could take information from the real world and then rebuild it into something that you could actually use. Little did we all know.
that just a few years later, AI technology powered by transformers was going to be a major facet of our lives. But it’s honestly been a bit underwhelming, if you ask me. Large language models have proven to be a technology that’s cool and fun, but not very practical. They can generate all sorts of output, from text to computer code to charts and graphs and tables and PowerPoint slides and crazy images and music and videos, but
Most of it isn’t very believable or even very useful. In fact, a lot of it just adds to the background noise in our daily lives. And a common criticism you’re starting to hear now about half-baked ideas is, did you have AI generate that? What’s more, research has shown that people tend to think less of those who use AI to create things than those who do all the hard work themselves. In this era of authenticity,
using AI as a shortcut that not only doesn’t impress people, but actually gives them a reason to doubt your credibility. Now, I don’t know about you, but I’m actually pretty proud of my ability to create things. When I write reports for my clients or script episodes for this podcast or come up with lectures for my students or write stories I want to share with readers, it’s all me, baby. 100 % pure, unadulterated Sean Jordan with maybe just the tiniest bit of help from a spell checker.
Could I have an AI platform create a reasonable facsimile of what I’d write or say? Absolutely, I’ve tried it myself. It was a little eerie the first time, if I’m being honest. But here’s the thing. When I started looking at it more closely, the output was saying things I would never say if you knew me.
It sounded like words someone else had written for me based on a vague idea of who I am, not the genuine article. And I’ve since noticed the same thing happens when I generate images or have a chat bot generate text or even when I have one of these tools try to analyze data. It looks good from far away, but once I start digging into it, I can see that the output isn’t valid. It isn’t reliable. And in the case of data where I can validate the findings,
isn’t even truthful. Because yes, AI platforms make up data and insert it into real data. And it’s really hard to distinguish fact from fiction without fact checking every single thing you see. Which leads me to ask, if the best these platforms can do is create a copy that’s only convincing from a quick glance, why are people making such a big deal about them? And the answer is, because we’d all like to have an easy button. OK, so.
Back in 2005, the office supply store chain Staples introduced this advertising campaign where they showed off this fictional gadget they called the Easy Button. It was a simple red button someone could press whenever they were doing something challenging like changing a diaper or breaking a horse or answering a math question or performing surgery. But the idea, of course, was that Staples was the easy button for ordering office supplies, fitting their tagline, That Was Easy. And the idea was so popular,
You can still buy easy buttons today to put on your desk if you really want to play up that audience nostalgia. There’s a real appeal to having an easy button in life. And when I do consumer research, what I hear about today isn’t staples, but Amazon, the online everything store that really has made it easy to pull out your phone, locate an item you want, and then order it to be delivered straight to your door within a day or
People like having that ability, but they also have complaints. For example, Amazon is full of knockoff products with make-believe names like ElecDare and Newt and Mosoter and Eilighen and IRAG and Jiminyar and a bunch of others I can’t even pronounce. Amazon also has really spotty customer service that can range from you can keep that product you’re trying to return and we’ll get you a full refund to
You need to provide us with documentary evidence from multiple angles that the Nintendo Switch 2 box you thought was supposed to contain a Nintendo Switch 2 was actually just a box filled with rocks, shredded newspapers, and old diapers. And it’s only getting worse. Because Amazon, like a lot of other companies, wants customers to interact with their automated systems before they even talk to a human being. AI fares really badly in customer service scenarios because it’s both sycophantic and completely unable to help people.
often because it’s not actually empowered to do anything, even if it promises that it will, all while wearing a blue blazer and a red tie. That’s the problem with having an easy button. You run the risk that it might give you a solution that’s not quite so easy to deal with if you’re unhappy with it. And I think about this a lot when it comes to generative AI platforms like ChatGPT, Copilot, Claude, Gemini, and even Amazon’s own Rufus.
They are so enticingly easy to use, but what happens when they get things wrong instead of right? And the consequences are more serious than just getting a bad Nintendo Switch 2 box from a retailer. For marketers, there are some serious considerations. For example, advertising is highly regulated and making false ad claims can result in very serious lawsuits and fines. Would you trust a large language model to follow the rules that even human beings can have a hard time following? I know I sure wouldn’t.
Also, who owns the copyright for AI output? There’s still no clear law or precedent. And the answer for now is no one really knows. This can be a big problem for an advertiser or marketer because that creative work is generally owned by the organization paying to create the content. And what about the general lack of originality? I mean, sure, AI can produce some really crazy stuff if you prompt it appropriately, but after a while,
starts to feel boring if everyone’s doing it. And there are only so many visual styles and ideas you can rip off before the audience feels like they’ve seen everything. And AI is really, really bad at creating things that are new or novel or even interesting because it can’t actually think or reason or innovate. It can only reformulate what it has in its training data. That’s a limitation that’s important to consider because it means AI is always backwards looking.
focused on what it already knows, not fresh new insights that move it forward. There are plenty of good applications for AI, but the key is that the people who are using it effectively are using it as a power tool to help them to start things, like a drill press putting screw holes into wood. Or they’re applying an extra layer of polish at the end on something they’ve already created, kind of like a power sander on something that’s already been carved and just needs to be smoothed out. That’s not using it as an easy button.
And I would recommend that’s the way we ought to be thinking about this technology as a tool. Because if we use it to start augmenting what we do, it’s no different from using spell check or Excel formulas or random number generators or anything else that helps us just be better at what we’re already doing. But if we try to use it to replace the work we do, the best we’re going to get is a bad facsimile of what we’ve already created. Well, I’m Sean in St. Louis, and this has been the Marketing Gateway.
See you next time.
Today’s plug is for Pac-Man. Now I mentioned him at the top of the episode, but I don’t know if people who don’t play video games realize just how incredibly awesome Pac-Man actually is. First of all, it’s been 45 years since Pac-Man started munching dots in arcades, and he’s still going strong with new video games and cartoon shows and merchandise and…
all the other trappings of a long and interesting career. Pac-Man might not be as visible as Mario or Sonic these days, but he’s still one of the most recognizable characters in popular culture, down to even having multiple songs named after him. Pac-Man has also been used to teach a lot of people how to code. One of the earliest lessons people learn in game development is how to make their own clone of the original Pac-Man game. It teaches them how to create a maze with walls that stop objects from moving through them.
how to create objects that disappear when a character moves over them, how to create a character you can move with a joystick, and of course, how to create enemies that use rudimentary AI to chase the player down. Beyond all of that, Pac-Man is a triumph of creativity and a truly interesting case study for how art and entertainment and interactivity can collide to create something that’s still interesting and enduring after five decades of existence. You can learn a lot.
from studying Pac-Man and the entire Pac family. You we’ve met his wife, his kids, his neighbors, and all sorts of other crazy characters over the last 45 years. And Pac-Man’s youngest, baby Pac-Man, even flirted with Pinball at one point. Pac himself became a professor at one point and ran a trivia video game. He’s also starred in puzzle games, adventure games, mobile games, fantasy games, and even feature films like Pixels. And did I mention Pac’s even been a character in games like
Super Smash Brothers Ultimate and Street Fighter X Tekken where he’s not only been able to duke it out with some of the biggest video game characters there are, but to show down with Ryu, Ken, Chun-Li, Kirby, Mario, and Sonic themselves. So if you haven’t played Pac-Man in a while, know, fire it up sometime and give it a try. It’s still a lot of fun and it might inspire you to come up with a wild and creative idea of your own.
Leave a Reply